Research On Underwater Target Detection Method Based On Deep Learning

Underwater target detection is one of the key technologies in the visual perception system of autonomous underwater robots.It uses image information collected by sensors to quickly and accurately identify target,and provide the necessary basis for subsequent target tracking and positioning.Traditional underwater target detection methods have limited feature extraction capabilities and generally low recognition rates.In this paper,deep learning methods are used to conduct in-depth research on underwater target detection.Underwater biological products(scallops,starfish,holothurian and echinus)are used as the detection objects.The research is carried out from two aspects of data enhancement and model design.Establish and train an underwater lightweight detection network model,and finally complete the physical experiment verification of the model on the Jetson TX2 platform.The main research work of thesis is as follows:(1)Research on underwater data enhancementAiming at the problems of poor image quality,small amount of data,and imbalance of target classes in aquatic product data sets,data augmentation and image processing are used to enhance data.Based on the improved deep convolutional generative adversarial network(DCGAN)to generate a large number of targets with fewer categories,balance and expand the training data set,and optimize the generator structure through feature fusion to improve the diversity of the generated images.Contrast limited adaptive histogram equalization algorithm(CLAHE)and Multi-scale retinex with color restoration algorithm(MSRCR)are used to solve the problems of low contrast and color cast of underwater images,and improve the quality of aquatic product data sets.(2)Research on underwater target detection algorithmIn view of the poor performance of traditional target detection algorithms in underwater environments,in-depth study of the performance of deep learning networks Faster R-CNN,SSD and YOLOv3 on underwater data.By comparing the experimental results,the YOLOv3 algorithm is used as a reference model for follow-up research.At the same time,the algorithm is used to detect the samples after data amplification and image enhancement.The results show that data enhancement can effectively improve the accuracy of underwater target detection algorithms.(3)Design of underwater lightweight detection networkIn allusion to problem of insufficient real-time performance of the YOLOv3 algorithm in embedded devices,a lightweight network model UW_YOLOv3 is designed for underwater target detection.In the backbone network,a deep convolutional separable network is used to replace the standard convolution,and parallel 1×1 convolution is introduced to increase the network width,and a residual structure is used to map the input to the output layer.The introduction of HRNet into the prediction network improves the traditional feature fusion method,so that the network has always maintained high-resolution feature representation capabilities.The experimental results show that the UW_YOLOv3 algorithm effectively improves the detection speed while maintaining a certain detection accuracy,and meets the needs of actual engineering applications.(4)Realization of real-time detection system for underwater targetsBased on the NVIDIA Jetson TX2 platform,an underwater real-time target detection system was built.Complete the deployment and testing of the lightweight algorithm on TX2,and use TensorRT to accelerate the forward calculation of the model.Based on PyQt5,the visual interactive interface of the underwater target detection system is designed,and the functions of image processing and target detection are integrated.Finally,the practical value of the software is demonstrated through the test.